xrtm-forecast 0.8.0

Institutional-grade modular engine for generative forecasting and agentic reasoning.

xrtm-forecast v0.7.0

The Runtime Engine for XRTM.

xrtm-forecast provides the agents, providers, topologies, and orchestration to build AI forecasting systems. It's a composable framework — import the pieces you need and wire them together.

Installation

pip install xrtm-forecast

Quick Start

from xrtm.forecast.kit.agents.specialists.analyst import ForecastingAnalyst
from xrtm.forecast.core.config.inference import OpenAIConfig
from xrtm.forecast.providers.inference.factory import ModelFactory

# Use any OpenAI-compatible endpoint


config = OpenAIConfig(
    model_id="your-model",
    base_url="$OPENAI_BASE_URL",
    api_key="your-api-key",
)
provider = ModelFactory.get_provider(config)
analyst = ForecastingAnalyst(model=provider, name="forecaster")

# Ask a forecasting question
from xrtm.data import ForecastQuestion
question = ForecastQuestion(
    id="fed-march-2025",
    title="Will the Fed raise rates in March 2025?",
)
forecast = await analyst.run(question)
print(f"Probability: {forecast.probability:.2f}")

What's Included

OpenAI-compatible Provider

One provider that works with any OpenAI-compatible API: any OpenAI-compatible endpoint. Set base_url to point anywhere.

Agents

• ForecastingAnalyst — Full forecast agent: structured prompts, causal reasoning traces, JSON output parsing
• LLMAgent — Base agent with tool retrieval, output parsing
• RoutingAgent — FAST/SMART tiered routing for cost optimization
• ToolAgent / GraphAgent — Wrap functions or sub-graphs as agents

Topologies

Composable multi-agent patterns in kit/topologies/:

• RecursiveConsensus — Parallel analysts → aggregate → supervisor → loop
• create_debate_graph() — Pro/Con/Judge debate pattern
• create_fanout_graph() — Parallel workers → aggregator

Orchestration

• Orchestrator — DAG state machine for workflow graphs
• AsyncRuntime — Chronos-aware async runtime

Infrastructure

• InferenceCache — SQLite cache for LLM responses (enabled by default)
• Retry with backoff — 2 retries on API errors
• LeakageGuardian — Temporal leakage prevention for backtests

XRTM Ecosystem

Part of the XRTM forecasting stack:

Package	Role	Version
xrtm-data	Schemas & question sources	0.3.0
xrtm-eval	Scoring (Brier, ECE, LogScore)	0.3.0
xrtm-forecast	Runtime engine (this package)	0.7.0
xrtm-train	Backtesting & optimization	0.3.0
xrtm	Product CLI	0.9.0

License

Apache 2.0